Individual mobility prediction by considering current traveling features and historical activity chain

Abstract

Individual mobility prediction forecasts traveling activities of an individual traveler, and has wide applications in location-based services, public health, and transportation planning. Whereas, it remains challenging due to the complexity and uncertainty of human mobility. Existing methods mainly consider spatiotemporal contexts in current traveling, but overlook those in historical trips, as well as relationships between traversed road intersections. These issues hinder the model from effectively capturing complex mobility patterns. To fill this gap, we propose a novel method that incorporates current traveling features and historical activity chain to predict the coordinates of traveling destination. Specifically, (1) we construct current traveling features by extracting real-time moving states, and represent spatiotemporal correlations between traversed road intersections using word embedding; (2) we learn travel intentions as a probability vector for each historical trip, and combine it with spatiotemporal features to construct historical activity chain; (3) we construct an individual mobility prediction model using Long Short-Term Memory (LSTM) network and spatiotemporal scoring mechanism, to capture short-term and long-term dependencies in current trip and historical activity chain, respectively. Experiments on 21,890 trajectories over the whole Year 2019 of 20 representatives selected from 1916 private car travelers in Shenzhen City, reveal the effectiveness of our model. It outperforms four baselines, Random Forest (RF), Distant Neighboring Dependencies (DND), Location Semantics and Location Importance (LSI)-LSTM, as well as Intersection Transfer Preference and Current Movement Mode (ITP-CMM), by approximately 10%-15% improvement in accuracy. In addition, we further explore the impact of historical activity chain length, and destination visiting frequency on prediction, as well as the relationship between predictability and eight mobility pattern features. This study benefits potential applications such as personalized location-based service recommendations and targeted advertising, and also provides implications for understanding human mobility.